continualai / continual-learning-baselines Goto Github PK

Continual learning baselines and strategies from popular papers, using Avalanche. We include EWC, SI, GEM, AGEM, LwF, iCarl, GDumb, and other strategies.

Home Page: https://avalanche.continualai.org/

License: MIT License

Python 99.76% Shell 0.24%

continual-learning lifelong-learning assessment experiments literature reproduction synaptic-intelligence elastic-weight-consolidation ewc gem

continual-learning-baselines's Introduction

Continual Learning Baselines

Avalanche Website | Avalanche Repository

This project provides a set of examples with popular continual learning strategies and baselines. You can easily run experiments to reproduce results from original paper or tweak the hyperparameters to get your own results. Sky is the limit!

To guarantee fair implementations, we rely on the Avalanche library, developed and maintained by ContinualAI. Feel free to check it out and support the project!

Experiments

The tables below describes all the experiments currently implemented in the experiments folder, along with their result. The tables are not meant to compare different methods but rather as a reference for their performance. Different methods may use slightly different setups (e.g., starting from a pre-trained model or from scratch), so it does not always make sense to compare them.

If an experiment reproduces exactly the results of a paper in terms of Performance (even if with different hyper-parameters), it is marked with ✅ on the Reproduced column. Otherwise, it is marked with ❌.
Avalanche means that we could not find any specific paper as reference and we used the performance of Avalanche obtained when the strategy was first add to the library.
If the Performance is much worse than the expected one, the bug tag is used in the Reproduced column.
Finally, the Reference column reports the expected performance, together with a link to the associated paper (if any). Note that the link does not always point to the paper which introduced the strategy, since it sometimes differs from the one we used to get the target performance.

ACC means the Average Accuracy on all experiences after training on the last experience.

First, we report the results for the non-online continual learning case (a.k.a. batch continual learning). Then, we report the results for the online continual learning case.

Batch Continual Learning (non-online)

Benchmarks	Strategy	Scenario	Performance	Reference	Reproduced
Permuted MNIST	Less-Forgetful Learning (LFL)	Domain-Incremental	ACC=0.88	ACC=0.88	✅ `Avalanche`
Permuted MNIST	Elastic Weight Consolidation (EWC)	Domain-Incremental	ACC=0.83	ACC=0.94	❌
Permuted MNIST	Synaptic Intelligence (SI)	Domain-Incremental	ACC=0.83	ACC=0.95	❌
Split CIFAR-100	LaMAML	Task-Incremental	ACC=0.70	ACC=0.70	✅
Split CIFAR-100	iCaRL	Class-Incremental	ACC=0.48	ACC=0.50	✅
Split CIFAR-100	Replay	Class-Incremental	ACC=0.32	ACC=0.32	✅ `Avalanche`
Split MNIST	RWalk	Task-Incremental	ACC=0.99	ACC=0.99	✅
Split MNIST	Synaptic Intelligence (SI)	Task-Incremental	ACC=0.97	ACC=0.97	✅
Split MNIST	GDumb	Class-Incremental	ACC=0.97	ACC=0.97	✅
Split MNIST	GSS_greedy	Class-Incremental	ACC=0.82	ACC=0.78	❌
Split MNIST	Generative Replay (GR)	Class-Incremental	ACC=0.75	ACC=0.75	✅
Split MNIST	Learning without Forgetting (LwF)	Class-Incremental	ACC=0.23	ACC=0.23	✅
Split Tiny ImageNet	LaMAML	Task-Incremental	ACC=0.54	ACC=0.66	❌
Split Tiny ImageNet	Learning without Forgetting (LwF)	Task-Incremental	ACC=0.44	ACC=0.44	✅
Split Tiny ImageNet	Memory Aware Synapses (MAS)	Task-Incremental	ACC=0.40	ACC=0.40	✅
Split Tiny ImageNet	PackNet	Task-Incremental	ACC=0.46	ACC=0.47 (Table 4 `SMALL`)	✅

Online Continual Learning

Benchmarks	Strategy	Scenario	Performance	Reference	Reproduced
CORe50	Deep Streaming LDA (DSLDA)	Class-Incremental	ACC=0.79	ACC=0.79	✅
Permuted MNIST	GEM	Domain-Incremental	ACC=0.80	ACC=0.83	✅
Split CIFAR-10	Online Replay	Class-Incremental	ACC=0.50	ACC=0.50	✅ `Avalanche`
Split CIFAR-10	ER-AML	Class-Incremental	ACC=0.47	ACC=0.47	✅
Split CIFAR-10	ER-ACE	Class-Incremental	ACC=0.45	ACC=0.52	✅
Split CIFAR-10	Supervised Contrastive Replay (SCR)	Class-Incremental	ACC=0.36	ACC=0.48	✅ `Avalanche`
Permuted MNIST	Average GEM (AGEM)	Domain-Incremental	ACC=0.81	ACC=0.81	✅
Split CIFAR-100	GEM	Task-Incremental	ACC=0.63	ACC=0.63	✅
Split CIFAR-100	Average GEM (AGEM)	Task-Incremental	ACC=0.62	ACC=0.62	✅
Split CIFAR-100	ER-ACE	Class-Incremental	ACC=0.24	ACC=0.25	✅
Split CIFAR-100	ER-AML	Class-Incremental	ACC=0.24	ACC=0.24	✅
Split CIFAR-100	Online Replay	Class-Incremental	ACC=0.21	ACC=0.21	✅ `Avalanche`
Split MNIST	CoPE	Class-Incremental	ACC=0.93	ACC=0.93	✅
Split MNIST	Online Replay	Class-Incremental	ACC=0.92	ACC=0.92	✅ `Avalanche`

Python dependencies for experiments

Outside Python standard library, the main packages required to run the experiments are PyTorch, Avalanche and Pandas.

Avalanche: The latest version of this repo requires the latest Avalanche version (from master branch): pip install git+https://github.com/ContinualAI/avalanche.git. The CL baselines repo is tagged with the supported Avalanche version (you can browse the tags to check out all the versions). You can install the corresponding Avalanche versions with pip install avalanche-lib==[version number], where [version number] is of the form 0.1.0. For some strategies (e.g., LaMAML) you may need to install Avalanche with extra packages, like pip install avalanche-lib[extra]. For more details on how to install Avalanche, please check out the complete guide here.
PyTorch: we recommend to follow the official guide.
Pandas: pip install pandas. Official guide.

Run experiments with Python

Place yourself into the project root folder.

Experiments can be run with a python script by simply importing the function from the experiments folder and executing it.
By default, experiments will run on GPU, when available.

The input argument to each experiment is an optional dictionary of parameters to be used in the experiments. If None, default parameters (taken from original paper) will be used.

from experiments.split_mnist import synaptic_intelligence_smnist  # select the experiment

 # can be None to use default parameters
custom_hyperparameters = {'si_lambda': 0.01, 'cuda': -1, 'seed': 3}

# run the experiment
result = synaptic_intelligence_smnist(custom_hyperparameters)

# dictionary of avalanche metrics
print(result)

Command line experiments

Place yourself into the project root folder.
You should add the project root folder to your PYTHONPATH.

For example, on Linux you can set it up globally:

export PYTHONPATH=${PYTHONPATH}:/path/to/continual-learning-baselines

or just for the current command:

PYTHONPATH=${PYTHONPATH}:/path/to/continual-learning-baselines command to be executed

You can run experiments directly through console with the default parameters.
Open the console and run the python file you want by specifying its path.

For example, to run Synaptic Intelligence on Split MNIST:

python experiments/split_mnist/synaptic_intelligence.py

To execute experiment with custom parameters, please refer to the previous section.

Run tests

Place yourself into the project root folder.

You can run all tests with

python -m unittest

or you can specify a test by providing the test name in the format tests.strategy_class_name.test_benchmarkname.

For example to run Synaptic Intelligence on Split MNIST you can run:

python -m unittest tests.SynapticIntelligence.test_smnist

Cite

If you used this repo you automatically used Avalanche, please remember to cite our reference paper published at the CLVision @ CVPR2021 workshop: "Avalanche: an End-to-End Library for Continual Learning". This will help us make Avalanche better known in the machine learning community, ultimately making it a better tool for everyone:

@InProceedings{lomonaco2021avalanche,
    title={Avalanche: an End-to-End Library for Continual Learning},
    author={Vincenzo Lomonaco and Lorenzo Pellegrini and Andrea Cossu and Antonio Carta and Gabriele Graffieti and Tyler L. Hayes and Matthias De Lange and Marc Masana and Jary Pomponi and Gido van de Ven and Martin Mundt and Qi She and Keiland Cooper and Jeremy Forest and Eden Belouadah and Simone Calderara and German I. Parisi and Fabio Cuzzolin and Andreas Tolias and Simone Scardapane and Luca Antiga and Subutai Amhad and Adrian Popescu and Christopher Kanan and Joost van de Weijer and Tinne Tuytelaars and Davide Bacciu and Davide Maltoni},
    booktitle={Proceedings of IEEE Conference on Computer Vision and Pattern Recognition},
    series={2nd Continual Learning in Computer Vision Workshop},
    year={2021}
}

Contribute to the project

We are always looking for new contributors willing to help us in the challenging mission of providing robust experiments to the community. Would you like to join us? The steps are easy!

Take a look at the opened issues and find yours
Fork this repo and write an experiment (see next section)
Submit a PR and receive support from the maintainers
Merge the PR, your contribution is now included in the project!

Write an experiment

Create the appropriate script into experiments/benchmark_folder. If the benchmark is not present, you can add one.
Fill the experiment.py file with your code, following the style of the other experiments. The script should return the metrics used by the related test.
Add to tests/target_results.csv the expected result for your experiment. You can add a number or a list of numbers.
Write the unit test in tests/strategy_folder/experiment.py. Follow the very simple structure of existing tests.
Update table in README.md.

Find the avalanche commit which produced a regression

Place yourself into the avalanche folder and make sure you are using the avalanche version from that repository in your python environment (it is usually enough to add /path/to/avalanche to your PYTHONPATH).
Use the gitbisect_test.sh (provided in this repository) in combination with git bisect to retrieve the avalanche commit introducing the regression.
git bisect start HEAD v0.1.0 -- # HEAD (current version) is bad, v0.1.0 is good
git bisect run /path/to/gitbisect_test.sh /path/to/continual-learning-baselines optional_test_name
git bisect reset
The gitbisect_test.sh script requires a mandatory parameter pointing to the continual-learning-baselines directory and an optional parameter specifying the path to a particular unittest (e.g., tests.EWC.test_pmnist). If the second parameter is not given, all the unit tests will be run.
The terminal output will tell you which commit introduced the bug
You can change the HEAD and v0.1.0 ref to any avalanche commit.